SgRNA targeting activity prediction method based on multi-granularity cross attention feature fusion

Abstract

The invention belongs to the technical field of gene editing and bioinformatics crossing, and particularly discloses an sgRNA targeting activity prediction method based on multi-granularity cross attention feature fusion, which comprises the following steps: step (1), acquiring a public data set: acquiring a high-throughput data set simultaneously containing sgRNA sequence information and a corresponding indel frequency (or editing efficiency) label, the data set at least comprises public data sets such as Sniper-Cas9, SpCas9, xCas9, HypaCas9, eSp-Cas9, CRISPRon (Clustered Regularly Interspaced Short Palindromic Repeats), HTCas9 (HyperCas9) and the like; according to the method, key features such as sequence information, DNA shape parameters, RNA secondary structures and chromatin accessibility are integrated into a four-branch deep learning architecture in a unified mode, a cross attention fusion mechanism combining coarse granularity and fine granularity is adopted for the first time, collaborative modeling and dynamic weighting are conducted on heterogeneous features, and the accuracy of sgRNA activity prediction is remarkably improved.

Images